Data-Driven Learning in an Incremental Grammar Framework

Overview Incremental processing of both syntax and semantics, both in parsing and generation, is of significant interest for modelling the human language capability, and for building systems which interact with it. Formal linguistics has made significant contributions to this; one example is the framework Dynamic Syntax, which provides an inherently word-by-word incremental grammatical framework. However, making this practical for computational models or systems involves building grammars with broad coverage on real data – a significant challenge. Here, we describe a method for inducing such a grammar from a corpus in which sentences are paired with semantic logical forms. By taking a probabilistic view, we hypothesise possible lexical entries – including entries for anaphoric elements – and learn a lexicon from their observed distributions without requiring annotation at the word level. The resulting grammar provides a resource for incremental semantic processing with good coverage, while learning grammatical constraints similar to a hand-crafted version.